P01
Computer Vision
Handwritten Digit Recognition using CNN
Computer Vision / Deep Learning
To design and train a Convolutional Neural Network (CNN) that classifies handwritten digit images (0-9) with high accuracy, demonstrating a foundational deep-learning image-classification pipeline.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, JupyterHub, NVIDIA A16 GPU
Dataset: MNIST Handwritten Digits Dataset (60,000 training / 10,000 test images)
- Loaded and normalized the MNIST dataset (pixel values scaled to 0-1).
- Built a CNN with two Conv2D + MaxPooling blocks followed by dense layers using Keras Sequential API.
- Compiled the model with the Adam optimizer and categorical cross-entropy loss.
- Trained for 10 epochs on the GPU node using the myenv conda environment via JupyterHub.
- Evaluated the model on the held-out test set and visualised the confusion matrix using Seaborn.
P02
Computer Vision
CIFAR-10 Image Classification using PyTorch
Computer Vision / Deep Learning
To build and train a custom Convolutional Neural Network in PyTorch capable of classifying natural images into 10 object categories (airplane, car, bird, cat, etc.).
Tools, dataset & methodology
Tools: Python, PyTorch, torchvision, JupyterHub, NVIDIA A16 GPU
Dataset: CIFAR-10 dataset (60,000 32x32 colour images across 10 classes)
- Loaded the CIFAR-10 dataset using torchvision.datasets and applied data augmentation (random crop, horizontal flip).
- Defined a 4-layer CNN with batch normalization and dropout for regularisation.
- Trained the network for 25 epochs using SGD with momentum on the GPU node.
- Tracked training/validation loss curves and computed per-class accuracy.
- Verified GPU utilisation using nvidia-smi during training.
P03
Computer Vision
Real-Time Object Detection using Intel OpenVINO
Computer Vision / Edge AI Optimisation
To demonstrate inference acceleration and deployment of a pre-trained object-detection model using the Intel OpenVINO Toolkit, optimised for real-time performance on Intel hardware.
Tools, dataset & methodology
Tools: Python, Intel OpenVINO Toolkit, OpenCV, pre-trained YOLO/SSD model
Dataset: COCO pre-trained object detection model, live webcam / sample video feed
- Converted a pre-trained TensorFlow/ONNX object detection model to OpenVINO Intermediate Representation (IR) format using the Model Optimizer.
- Loaded the IR model using the OpenVINO Inference Engine on the compute node.
- Ran inference on a sample video stream, drawing bounding boxes and confidence scores using OpenCV.
- Benchmarked inference latency before and after OpenVINO optimisation.
P05
NLP
Email Spam Classification using Machine Learning
Natural Language Processing / Classical ML
To build a text classification pipeline that distinguishes spam from legitimate (ham) messages, and to compare training performance with and without Intel's oneAPI-accelerated Scikit-learn extension.
Tools, dataset & methodology
Tools: Python, Scikit-learn, Pandas, Seaborn, Intel Extension for Scikit-learn (oneAPI AI Analytics Toolkit)
Dataset: SMS/Email Spam Collection dataset (5,500+ labelled messages)
- Cleaned and vectorised message text using TF-IDF.
- Trained a Multinomial Naive Bayes classifier and a Logistic Regression classifier for comparison.
- Patched Scikit-learn with the Intel Extension for Scikit-learn (available via the Intel AI Analytics Toolkit) to benchmark CPU training speed-up.
- Evaluated both models using accuracy, precision, and recall; visualised results with a Seaborn heatmap.
P06
Machine Learning
House Price Prediction using Regression Models
Machine Learning / Regression
To predict housing prices using regression techniques and compare the performance of Linear Regression against ensemble methods.
Tools, dataset & methodology
Tools: Python, Pandas, Scikit-learn, Seaborn, JupyterHub
Dataset: Boston/California Housing dataset (numeric features: rooms, location, income, etc.)
- Performed exploratory data analysis and correlation heatmaps using Pandas and Seaborn.
- Trained Linear Regression, Ridge Regression, and Random Forest Regressor models.
- Tuned hyperparameters using GridSearchCV with 5-fold cross-validation.
- Compared models using RMSE and R² score on the held-out test set.
P07
Computer Vision
Plant Leaf Disease Detection using Transfer Learning
Computer Vision / Agri-Tech AI
To detect and classify plant leaf diseases from images using transfer learning on a pre-trained MobileNetV2 model, targeting an agriculture-focused, resource-efficient AI use case.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras (MobileNetV2), NVIDIA A16 GPU
Dataset: PlantVillage dataset (leaf images across healthy and diseased categories)
- Loaded the PlantVillage dataset and applied image augmentation (rotation, zoom, flip).
- Used MobileNetV2 (pre-trained on ImageNet) as a frozen feature extractor with a custom classification head.
- Fine-tuned the top layers for 12 epochs on the GPU node.
- Evaluated per-class precision/recall and visualised sample predictions with Grad-CAM-style heatmaps.
P08
Computer Vision
Face Mask Detection using CNN and OpenCV
Computer Vision
To build a real-time face mask detection system combining a CNN classifier with OpenCV-based face detection, relevant to public-health compliance monitoring.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, OpenCV, JupyterHub
Dataset: Face Mask Detection dataset (with-mask / without-mask labelled images)
- Trained a CNN classifier to distinguish 'mask' vs 'no-mask' cropped face images.
- Integrated OpenCV's Haar Cascade / DNN face detector to locate faces in a video frame.
- Combined face detection with the trained classifier for real-time annotated video output.
- Measured detection accuracy and end-to-end frame processing speed.
P09
Computer Vision
Traffic Sign Recognition using CNN
Computer Vision / Intelligent Transportation
To classify traffic sign images into 43 categories, a foundational task for autonomous-driving and driver-assistance systems.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, NVIDIA A16 GPU
Dataset: German Traffic Sign Recognition Benchmark (GTSRB) — 43 sign classes
- Preprocessed and normalised GTSRB images to a fixed input size of 32x32.
- Built a deeper CNN with 3 convolutional blocks and dropout regularisation.
- Trained for 20 epochs on the GPU node with a learning-rate scheduler.
- Evaluated per-class accuracy, noting classes with visually similar signs.
P10
NLP
Fake News Detection using NLP
Natural Language Processing
To classify news articles as fake or real using classical NLP feature extraction and a Logistic Regression classifier, addressing a socially relevant misinformation-detection use case.
Tools, dataset & methodology
Tools: Python, Scikit-learn, Pandas, TF-IDF
Dataset: Fake and Real News dataset (labelled news articles)
- Cleaned article text (stop-word removal, lowercasing, punctuation stripping).
- Extracted TF-IDF features with a vocabulary of the top 20,000 terms.
- Trained and compared Logistic Regression and Passive-Aggressive classifiers.
- Evaluated using accuracy, confusion matrix, and manually inspected misclassified samples.
P11
Machine Learning
Customer Churn Prediction using Ensemble Learning
Machine Learning / Business Analytics
To predict whether a customer will churn (discontinue service) based on usage and account features, supporting a business-analytics style application of AI.
Tools, dataset & methodology
Tools: Python, Pandas, Scikit-learn, Seaborn
Dataset: Telecom Customer Churn dataset (7,000+ customer records)
- Performed data cleaning and encoding of categorical features (contract type, payment method, etc.).
- Handled class imbalance using SMOTE oversampling.
- Trained a Random Forest and a Gradient Boosting classifier, comparing performance.
- Analysed feature importance to identify key churn drivers.
P12
Deep Learning
Human Activity Recognition using LSTM
Time-Series / Deep Learning
To classify human physical activities (walking, sitting, standing, climbing stairs) from wearable sensor time-series data using an LSTM network.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, Pandas
Dataset: UCI HAR dataset (smartphone accelerometer/gyroscope sensor readings)
- Segmented raw accelerometer/gyroscope signals into fixed-length windows.
- Built a stacked LSTM network to model temporal dependencies in the sensor readings.
- Trained for 30 epochs with a validation split, monitoring for overfitting.
- Visualised the confusion matrix across the 6 activity classes.
P13
Deep Learning
Speech Emotion Recognition using CNN
Audio Processing / Deep Learning
To recognise emotional states (happy, sad, angry, neutral, etc.) from speech audio using MFCC feature extraction and a CNN classifier.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, Librosa (audio feature extraction)
Dataset: RAVDESS Speech Emotion dataset (8 emotion classes)
- Extracted Mel-Frequency Cepstral Coefficients (MFCCs) from each audio clip using Librosa.
- Reshaped MFCC features into 2D 'image-like' input for a CNN.
- Trained the CNN for 25 epochs on the GPU node.
- Evaluated per-emotion accuracy and identified commonly confused emotion pairs.
P14
Machine Learning
Diabetes Prediction using Classification Models
Machine Learning / Healthcare Analytics
To predict the likelihood of diabetes in patients based on diagnostic measurements, illustrating an AI application in preliminary healthcare screening.
Tools, dataset & methodology
Tools: Python, Scikit-learn, Pandas, Seaborn
Dataset: Pima Indians Diabetes dataset (768 patient records, 8 clinical features)
- Handled missing/zero values in clinical features (glucose, BMI, insulin) via median imputation.
- Standardised features and trained Logistic Regression, SVM, and Random Forest classifiers.
- Compared models using accuracy, sensitivity (recall), and specificity — prioritising recall given the healthcare context.
- Visualised feature correlations using a Seaborn heatmap.
P15
NLP
AI Chatbot using Intent Classification
Natural Language Processing / Conversational AI
To build a rule-assisted conversational chatbot that classifies user input into predefined intents and responds appropriately, demonstrating a practical NLP application.
Tools, dataset & methodology
Tools: Python, TensorFlow, Keras, NLTK, JSON-based intent dataset
Dataset: Custom intent dataset (greetings, FAQs, college-info style intents, ~150 patterns)
- Defined an intents.json file with sample patterns and responses across ~15 intent categories.
- Tokenised and lemmatised patterns using NLTK, converting text to a bag-of-words representation.
- Trained a feed-forward neural network (Dense layers) in Keras to classify user input into an intent.
- Built a simple loop-based chat interface that predicts intent and returns a matching response.
